IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 7, NO. 1, MARCH 2015 39 Suspenser: A Story Generation System for Suspense Yun-Gyung Cheong, Member, IEEE, and R. Michael Young, Senior Member, IEEE

Abstract—Interactive storytelling has been receiving a growing when they are anticipating the occurrence of some event and attention from AI and game communities and a number of compu- are uncertain about the event's outcome [13]–[15]. Our work tational approaches have shown promises in generating stories for focuses on the class of suspense associated with the perceived games. However, there has been little research on stories evoking specific cognitive and affective responses. The goal of the work likelihood of undesirable outcomes over preferred outcomes. we describe here is to develop a system that produces a narra- More specifically, implicit in much of the work we present here tive designed specifically to arouse suspense from its reader. Our is the notion articulated by GerrigandBernardo[16]inwhich approach attempts to create stories that manipulate the reader's they view a story's audience as active problem-solvers. In their suspense level by elaborating on the story structure that can influ- model of narrative comprehension, a reader's feeling of sus- ence the reader's narrative comprehension at a specific point in her pense is affected by the number of potential solutions she can reading. Adapting theories developed by cognitive psychologists, our approach uses a plan-based model of narrative comprehension anticipate for the dilemma faced by the protagonist. to determine the final content of the story in order to manipulate To generate suspenseful stories, we set out a basic approach the reader's suspense. In this paper, we describe our system im- built on a tripartite model, adapted from narrative theory, that plementation and empirical evaluations to test the efficacy of this involves the following elements: the fabula,thesjuzhet,andthe system. discourse [17]–[20]. A fabula is a story world that includes all Index Terms—Cognitive model, discourse planning, narrative the events, characters, and situations in a story. A sjuzhet is a generation, suspense. series of events selected from the fabula and an ordering over those events indicating the order in which they are to be pre- sented to readers. In our approach, the fabula and sjuzhet are I. INTRODUCTION represented as plan structures. The final layer, the discourse, ARRATIVE has long been associated with games. While can be thought of as the medium of presentation itself (e.g., N most games employ linear narrative as a back story only text, film). In this paper, we present Suspenser, a framework that to engage the gamer with gameplay, classical adventure games determines narrative contents (i.e., sjuzhet) from a given story and some recent action games such as Heavy Rain [1], L.A. world (i.e., fabula) intended to evoke suspense on the part of the Noire [2], Bioshock Infinite [3], and Fallout: New Vegas [4] reader. The rest of this paper is organized as follows. Section II are designed to offer the player with high narrative experience reviews relevant work in narrative psychology and AI. Section as well. A number of story generation systems and AI tech- III describes our tripartite model for story generation. Section niques have been developed for creating highly interactive sto- IV details the Suspenser framework, followed by our evalua- ries [5]–[9]. And yet, there has been little attention to affective tion that assesses the performance of Suspenser in Section V. properties of narrative, which are, in fact, fundamental to its ap- The last chapter concludes with a discussion of the limitations preciation by the user. The feeling of suspense, surprise, and of our system and the plan for future work. curiosity is generally expectedinreadingandviewingnarrative forms [10]–[12]. Among these emotions, suspense has drawn much attention from psychologists due to its significant impact II. RELATED WORK on the reader's enjoyment. In the empirical study conducted by Theories in psychology and cognitive science view the sus- Brewer and Lichtenstein [10], the participants reported that sus- pense in the reader's mind as related to diverse narrative el- pense is cardinal for discerningastoryfromamereseriesof ements. The reader's disposition toward protagonists or out- events and expressed high satisfaction about their narrative ex- comes is a primary condition for her to feel suspense [13], [21]; periences when suspenseful events were presented in the stories. the reader should be able to anticipate the potential for con- Our approach addresses the problem of creating suspense in flicting outcomes of an event [12]. The reader's suspense would narrative, which keeps the user engaged in the various plots, increase when she perceives likelihood of undesirable outcomes giving them high entertainment value. We view suspense as the relative to preferred outcomes [14], [15], [22]–[26]. Prolonging feeling of excitement or anxiety that audience members feel the length of time the reader anticipates an event or outcome that is potentially harmful to a protagonist can also heighten her sus- Manuscript received July 10, 2013; revised February 03, 2014; accepted May pense [26]. Discrepancies between the knowledge of the story 06, 2014. Date of publication May 14, 2014; date of current version March 13, world held by characters and that held by the reader can be 2015. This work was supported in part by Award 0092586 and Award 0414722, and the EU FP7 ICT Project SIREN (Project: 258453). used as a device to induce the reader's suspense [11], [12], [23]. Y.-G. Cheong is with the College of ICE, Sungkyunkwan University, South In the well-known emotion model OCC [13], suspense could Korea (e-mail: [email protected]). be regarded as a specific type of prospect-based emotion,one R. M. Young is with the Department of Computer Science, North Carolina State University, Raleigh, NC 27695 USA (e-mail: [email protected]). which is evoked when an individual anticipates the occurrence Digital Object Identifier 10.1109/TCIAIG.2014.2323894 of events whose outcome is uncertain. Their account of suspense

1943-068X © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information. 40 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 7, NO. 1, MARCH 2015 involves hope and fear; a person hopes his favorable conse- quence will be realized while he fears the occurrence of undesir- able consequences. Gerrig and Bernardo [16] hypothesize that a reader's feeling of suspense is affected by the number of poten- tial solutions available for the protagonist. Under this model, an audience will feel an increased measure of suspense as the number of options for the protagonist's successful outcome(s) decreases. To confirm this hypothesis, Gerrig and Bernardo per- formed a number of experiments with human subjects. The sub- jects were provided different text versions of a story where a protagonist is in danger and tries to escape. The various ver- sions of the story differed in the number of solutions available to the protagonist. After reading the text, subjects were asked to rate their estimation of the likelihood of the protagonist's escape as well as their suspense levels. The data from the experiments showed that the readers reported higher suspense as possible so- lutions were eliminated. Quite a number of interactive storytelling systems have Fig. 1. Example fabula plan structure, illustrating a father getting a toy for his seven-year old son Ben as a Christmas gift. In the diagram, temporal ordering been implemented so far [5]–[9], [27]–[31]. In particular, constraints ( ) proceeds from the top to the bottom. Rectangles represent a MINSTREL [31], a case-based approach to modeling human series of plan steps ( ), with each action's preconditions enumerated above its creativity, is most comparable to our approach. At the core rectangle. An arrow between two actions indicates a causal relationship that holds between the two ( ), meaning that the effects of the action at the starting of the MINSTREL design is a transform-recall-adapt process. point of the arrow establishes a precondition for the action at the arrow's end On receiving a problem specification as input, MINSTREL point. Given an initial state (i.e., the father is poor and he has a ring), this plan retrieves a case from its memory that is similar to the problem is constructed automatically to achieve the goal (has Ben toy). The condition of S3:Move action is true under the closed-world assumption: given as input. If the case is identical to the problem, the a condition not specified in the initial state is assumed to be false. The plan can original solution is used. If the case differs from the input be described in text as: “A poor father traded his wedding ring for the toy that problem, the original solution is modified. As an end-to-end his son Ben wants to have. He then put the toy under their Christmas tree. The next day Ben walked to the tree and found the toy that his father left.” system, MINSTREL extensively attempts to solve multiple story-generation issues such as themes, coherency, characteri- zation, tragedy, suspense, and foreshadowing. In MINSTREL suspense is created by fabula-level generation, relying on the course generator. The fabula is represented as a plan data struc- psychological evidence that readers feel more suspense when ture, as in Definition 1, created in response to a story request they strongly care about the character [13], [21] and when specifying the initial and goal states of the story world. To gen- the presentation of a significant outcome is prolonged [26]. In erate the fabula plan, we use an implementation of the hierar- our approach suspense is created by sjuzhet-level generation, chical, partial-order causal link planner Longbow [34], [35]1, focusing on selecting relevant contents when a fabula is given which is discussed in Section IV-A. A fabula (illustrated in as an input. Fig. 1) is sent as input to the second component in the pipeline, The advantage of using Suspenser is that it can work with Suspenser in our work. an already existing fabula generator such as IPOCL [32] which Definition 1 (Fabula): A fabula F is a tuple generates a story that is coherent with each character's intent or where is a series of plan steps, is a set of binding con- CPOCL [33] that plans a story containing conflict. The draw- straints on the variables in steps in , is a set of temporal back of Suspenser is that it cannot add a new story event that ordering constraints over steps in , is a list of causal links is not present in the fabula. For instance, imagine a scene that between steps in . Each element of is a step representing an an agent is climbing on a high-rise building with special gloves instantiation of a plan operator contained in a plan library. A on. To make the situation more suspenseful, one of his glove plan operator is a tuple where is a unique gadget runs out of battery and let him fall. Suspenser cannot name for the operator, is a set of preconditions, terms repre- add such an event that endangers the protagonist at the sjuzhet senting just those conditions that must hold for operator to be level without consulting with the fabula generator. On the other able to occur, and is a set of effects, terms denoting just those hand, Minstrel can create such an event to heighten suspense. conditions that change as a result of the action's successful exe- However, Minstrel needs to have a case that instructs how to cution. A causal link between two steps and , indicated by generate necessary events in a particular situation. the notation indicates that establishes as an effect some condition that is used to establish the precondition of step . III. A TRIPARTITE ARCHITECTURE OF STORY GENERATION From this fabula, Suspenser constructs a sjuzhet as defined in Definition 2. Upon receiving the story structure from Suspenser, Following the recent research in narratology claiming the the discourse generator produces surface structure (i.e., text, an- benefits of three-layer model in story analysis [17]–[20], our imation). The discourse generator in our current system uses a architecture is designed as a three-stage pipelined architecture which consists of fabula generator, sjuzhet generator, and dis- 1In this work, hierarchical planning functionality was not used. CHEONG AND YOUNG: SUSPENSER: A STORY GENERATION SYSTEM FOR SUSPENSE 41 template-based approach which maps a plan step into a single sentence. Suspenser serves as a sjuzhet generator in our work. Definition 2 (Sjuzhet): A sjuzhet is a tuple where is a fabula, is a subset of the plan steps of to be presented to the user, and is a set of temporal ordering con- straints over steps in .When is empty, uses the ordering and binding information of . We sketch selected narrative-planning models that can serve as a component for each layer. The first layer, the fabula gener- ator, can be replaced by a number of systems that enrich the story world by populating story characters and events. Riedl and Young [32] have developed Fabulist, a story generation system using an Intent-driven Partial-Order Causal Link planner (IPOCL). Fabulist ensures that the actions in the generated story are consistent with characters' intention. In this manner, Fabulist can maintain the balance between plot coherency and character believability. Further, an extension of the IPOCL algorithm is made by [33] to develop a story planner which can create con- Fig. 2. Suspenser architecture. flicts in narrative. Unlike conventional planners where threat- ened causal links are strictly prohibited, Conflict Partial-Order usedtoconveythestoryupto to a reader for the suspense level. Causal Link (CPOCL) allows them under the condition that Suspenser consists of three components: the skeleton builder, those events threatening causal links are not executed. Cavazza the suspense creator, and the reader model (see Fig. 2). The et al. [5] have developed a story generation system that builds a skeleton builder identifies a series of kernels [42], [43] in the storyline by modeling interactions between autonomous agents story —important events in a story that cannot be eliminated using Hierarchical Task Network (HTN) planning techniques. without harming the story's understandability. Those events that HTN planning represents a plan as a collection of possible sub- are not included in the skeleton are defined as Satellites,events tasks to achieve a higher level goal. The system has been further that enrich or elaborate upon the kernels and can be omitted extended to generate stories that can express the emotions felt without damaging the storyline. The reader model takes the se- by characters through the use of emotional operators which con- quence of kernels and checks them for coherency, essentially tain mental states as preconditions and effects [36], [37]. modeling a human reader's process of story comprehension to As a sjuzhet generator there are some systems that construct determine what mental model of the story a reader would form, an output story by manipulating story events and temporal or- as described later. The sequence is then passed to the suspense derings. [38] describes a reasoning-based approach that assem- creator that again uses the reader model to predict which story bles two distinct versions of one story that share climax into a elements from the sjuzhet can serve to contribute to manipulate single twisted story. Bae and Young [39] have investigated the suspense. use for telling events out of chronological sequence focusing on telling future events earlier than their occurrences ( fore- A. The Reader Model shadowing) or later than their occurrences ( flashback) aiming The reader model—acting as a proxy for an individual at reader's surprise arousal. reader's comprehension processes—contains four elements: Discourse generation relates to different media. For narra- a reasoning algorithm, a limit on reasoning capacity, a plan tive prose generation, Callaway and Lester [40] have developed library indicating the reader's background knowledge, and the AUTHOR system that performs narrative segmentation and a set of preferences that model a reader's preferred types of chooses appropriate pronoun for a concept, and makes lexical story elements. For simulating the human reader's reasoning choices that give variation to repetitive expressions. Their ex- process in this paper, we use Crossbow, a C# implementation perimental study supports that making decisions taking into ac- of the hierarchical partial-order causal link planner Longbow count the discourse history has a significant impact on the nar- [34], [35]. Prior work has provided strong evidence that types rative prose quality. In 3D-virtual environments, Darshak, the of human task reasoning are closely related to partial-order camera planning system [41] takes as input the story events planning algorithms [44] and that refinement search [45], the along with communicative goals and outputs a discourse plan type of plan construction process performed by Crossbow, integrating the camera directive actions into the input story plan. can be used as an effective model of human plan reasoning The selection of camera actions are made to ensure that the processes [46]. As a form of reasoning limit, an integer counter given communicative goals are achieved. is used to constrain the number of nodes to be searched during the planning process. To represent the reader's knowledge,we IV. A COMPUTATIONAL MODEL OF SJUZHET GENERATION use a plan library, consisting of a set of plan operators; each FOR SUSPENSE operator consists of a unique name, a set of preconditions and Suspenser takes four elements as input: a fabula, a suspense effects, and a set of variables to be instantiated in the planning level (either high or low), a goal, and a given point in the story process. The preconditions of an action in a plan represent just plan. Then Suspenser determines the sjuzhet, the content to be those conditions that must hold for the action to be able to 42 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 7, NO. 1, MARCH 2015

Fig. 3. Plan space graph modeling the reader's forward inference process. The root node #1 represents a partial description of a story's plot as a partial plan which consists of move and find actions along with causal links. Each arc between two nodes indicates an inference by the reader that reconstructs a single aspect of the story's missing detail as a plan structure. This inference process here leads to the construction of three different complete plans (#2, #4, #5). The node #2 is a complete plan that refines the parent node by adding the put action bound with Santa as the agent of the action, while the node #3 denotes a partial plan refining the parent node by adding the same put action bound with Dad as the agent. To satisfy the open precondition of (has Dad toy), two complete plans in the node #4 and the node #5 are generated by adding different actions trade and steal respectively. The plan in the node #4 corresponds to the fabula plan in Fig. 1. happen while the set of effects denotes just those conditions a single flaw from the parent node. When the flaw is an open that change by the action's successful execution. The preference precondition, a causal link is established in the new node's plan function models the reader's heuristic function used during the from either an existing step in the plan or an instantiated oper- refinement search performed as a part of the planning process. If ator in the plan library which has an effect that can be unified the reasoned story has no break in causal relationship between with the precondition; in the second case, the instantiated step one event to the next, then it will be understood as coherent. is also added to the parent plan. When the flaw is a threatened When a partial description of a story is given, the reader may causal link, a temporal constraint (i.e., either demotion or pro- infer a complete story by filling the gaps to understand it as a co- motion) to resolve the threat is added or binding constraints are herent story to achieve the protagonist's mission. In refinement added to separate the threat involved steps so that no conflicts search [45], the planning process is modeled as a search through arise. Fig. 3 shows a plan space resulting from expanding par- a space of plans which is represented as a directed acyclic graph tial plan #1 into three different complete plans (#2, #4, #5) by of partial plan nodes. An arc from one node to another indicates refining open precondition flaws in parent nodes. that the plan at the child node is a version of the plan from the parent node with a single plan flaw repaired. In our approach, the root node of the graph is a partial plan taken from the skeleton B. The Skeleton Builder builder or the suspense creator. Interior nodes in the plan graph represent partial plans with flaws, while leaf nodes of the graph Suspenser's primary task, selecting which story elements to are either complete plans without flaws or plans with flaws that tell, eliminates some of the story plan's elements from the dis- cannot be repairable due to inconsistency in the plan. In plans course describing the story. As more and more elements are ex- created by Crossbow, a flaw is either a precondition of some cluded from the discourse, however, the resulting gaps in the step that has not been established by prior step in the plan, or a plan may make the underlying fabula difficult to identify. For causal link that is threatened (i.e., undone) by the effect of some example, a story that leaves out the event of Cinderella losing other step in the plan. When adding a new node to the plan space her shoe would not be readily recognized as the well-known ver- graph, Crossbow creates the child node based on the repair of sion of the Cinderella story. To maintain the identity of the input CHEONG AND YOUNG: SUSPENSER: A STORY GENERATION SYSTEM FOR SUSPENSE 43 story, our approach selects a set core events of the fabula, ker- from an action to a goal as the number of actions interposed be- nels, for inclusion in the sjuhzhet using techniques that exploit tween them in a subgoal hierarchy constructed in the reader's results from narrative comprehension studies. mind. In our system, the distance from an action to the goal is The skeleton generator rates the importance of each event defined as the minimum number of causal links that relate an based on a method devised by Trabasso and Sperry [47] for ex- action to the goal in a plan. This condition requires the causal tracting important actions that are likely to be included in the distance magnitude ordering of action-goal pairs to be retained story recall. To determine an individual story event's quantita- in pair magnitudes, as well. For example, in tive importance, their approach counts the number of causal re- Fig. 1 the causal distance between the action S4:Find and goal lationships with other events. Further, they measure each event's step is 1, and the distance between the action S2:Put and the qualitative importance by analyzing its role in the causal chains. goal is 2, which makes the distance of the action S4-goal pair Causal chains are a series of events in the story that are causally nearer than that of the action S2-goal pair. Therefore, a function related. Causal chains contain actions that can be characterized suitable for would yield a value for the ac- as either opening events, closing events, or continuing events. tion S4-goal pair smaller than that for the S2-goal pair. Opening events introduce characters and the setting and initiate Once each event's importance is computed based on Heuristic the story. Closing events determine whether the protagonists' Function 1, the skeleton generator selects the most important main goals are achieved or not. Continuing events causally con- events in the story as kernels. The integer value of , a story's nect opening events to the closing events in the story. desired length, in our experiments was chosen by a human story Drawing on their approach, calculating the importance of writer. Once a skeleton is built, the coherency evaluator tests each event in a plan was formalised as shown in Heuristic whether the skeleton is coherent from the reader's perspective as Function 1. Our system approximates an event's quantitative im- an integral story using an algorithm which is a cycle composed portance by counting the number of its incoming and outgoing of two phases. The first step uses the reasoning algorithm in the causal links. For measuring an event's qualitative importance, reader model to find complete plans to achieve the protagonist's we define importance categories: opening-act, closing-act, goals which are consistent with the skeleton candidate. If such motivated-act, and contiuing-act types. Opening-acts are the a plan is found, the story skeleton is coherent and the program first actions in the story —those that connect propositions from exits. Otherwise, a satellite event in the fabula with the highest the initial state to later events in the text; Closing-acts are the importance value is chosen and added to the candidate. Then, last actions that occur in the story; motivated-acts are the plan the first phase begins again. steps that are in causal relationship with preconditions of the goal state; Continuing-acts are a default type since every step C. The Suspense Creator in a complete plan is causally related. 1) Heuristic Function 1 (Importance of an Action): The suspense creator takes as input the story skeleton and im- returns the importance of an action in plan portance values of satellites (see Fig. 2). The central task of the suspense creator is to find a series of additional plan steps from a portion of the story preceding the step , which enables the reader to infer the minimum number of solutions. The sus- pense creator consists of two subcomponents: the structure or- where returns the number of incoming causal links to ganizer and the suspense measurer. The structure organizer uses coming from steps of the plan other than the initial step, the skeleton as the initial contents of the sjuzhet and then it aug- returns the number of incoming causal links to from ments the skeleton with additional story elements that would the plan's initial step, returns the number of 's out- result in heightening suspense level. The suspense measurer going causal links, returns the causal-chain value of checks uncertainty and returns an estimated level of suspense in the plan, and returns a value associated when a sjuzhet is given as input. The interactions between these with the causal distance between the step and the goal step of two components are controlled by the suspense creator, as de- the plan . scribed in Algorithm 1. All scaling factors in the function ( , , , )arecon- The algorithm initializes the content of the sjuzhet with the strained to be real numbers no less than 0. In the formula, the skeleton given from skeleton builder. In order to find supple- causal-chain value of an event (that returns) is deter- mental story event contributing to suspense arousal, we define mined by the event's act type. Important categories (i.e., moti- the term potential suspense that refers to the amount of each vated-acts) are assigned high integer values to give increased event's contribution to the suspense level increase (Heuristic likelihood for those acts to be included in the skeleton. Less im- Function 3). Our system selects , the action with the greatest portant categories (i.e., continuing-act) are assigned low integer potential suspense from the set of satellite events . If the poten- values. The assigned values are determined empirically through tial suspense of is lower than a predefined threshold, then the informal experiments. When a step has multiple act types, the program returns as the selected story events and exits. Oth- type with the highest act value will be chosen. The definition of erwise, the system chooses an action with the lowest impor- the normalization function is informed by a tance in and computes its potential suspense value . psychological distance effect, which indicates that an action in If the value is lower than the potential suspense of the new event an episode is more readily comprehended when it is nearer to ( ), the system builds a new partial plan that re- the episode goal [48]. Foss and Bower [48] define the distance places the event with the event from the set . 44 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 7, NO. 1, MARCH 2015

A modification to the algorithm can create a story to elicit Algorithm 1: The Suspense Creator procedure a low suspense level from the reader. The algorithm selects an action with the lowest potential suspense and replaces the input : as a fabula plan containing a set of steps, a set of action in with the highest importance value if the current causal links, and a set of temporal constraints; suspense level is lowered by replacing with . as the step where the reader's suspense is measured; The following sections detail the two subcomponents of the suspense creator. When a sjuzhet is given, the suspense mea- as the portion of the skeleton preceding ; surer checks uncertainty condition and estimates suspense level as the set of importance weight for the events in ; based on Heuristic Function 2. The structure organizer con- structs a sjuzhet by augmenting story events that are likely to as the protagonist's goal; increase suspense based on Heuristic Function 3, which makes as the reader's resource bound; use of the values calculated by Heuristic Function 4. 1) The Suspense Measurer: AsdiscussedinSectionII,one as a planning algorithm; critical condition for a reader to feel suspense is to keep her as the plan library as the reader's knowledge; uncertain about the outcome of a significant event. When the reader is certain about the negative outcome, she may feel dis- output: as the content selected for suspense appointment or sadness rather than suspense [15]. To meet the uncertainty condition of suspense, the suspense measurer first ; checks if the reader model would be uncertain about the goal (events in the portion of preceding )– ; state using the planning space. In logical terms, an agent is un- certain about a proposition when the agent makes inferences, while is not empty do leading to the possibilities of being true and false [49]. The plan- an action in with the highest ; ning space represents the reader's reasoning and an inference corresponds to a path from the root node to a terminal node in if then the planning space. Therefore the reader model returns certainty Return ; when the planning space contains either only complete plans (absolute success) or only failed plans (absolute failure). There- end fore, the reader model returns uncertainty about the goal state ; when the planning space has both successful plans and failed plans. The reader model also returns uncertainty when the plan- an action in with the lowest importance in ; ning space exceeds the searching limit, based on the rationale if then that successful plans can be found if unlimited resources are given. ; When the uncertainty is ensured, the suspense measurer esti- mates the reader's suspense level, following the narrative com- ; prehension model by Gerrig and Bernardo [16]. The reader's if passes Uncertainty Checking then suspense increases as the number of options for the protagonist's successful outcome(s) decreases. Therefore, Heuristic Function if 2 computes the reader's suspense level as the inverse of the then number of planned solutions that achieve the protagonist's goal ; using her reasoning algorithm and her plan library within her reasoning limit. The function sets a minimum level of suspense end when no usable solutions are found in her plan space, as is sup- end ported by psychological research. 2) Heuristic Function 2 (Level of Suspense): end returns the level of suspense when end . Otherwise, it returns 0.

The function returns a partial plan structure which consists of plan steps and their causal links and temporal or- derings coherent with . Then the system queries the suspense where is a set of literals representing the goal of a narrative's measurer for the suspense levels of the new sjuzhet and protagonist, is a partial plan, is a plan library, is a plan- of the current sjuzhet (computed by Heuristic Function ning algorithm, is an integer representing a reasoning bound, 2). If the suspense level from is higher than that of and returns the number of paths to make , the system updates the current with the events in true with given and . . This process continues until there is no candidate is 3) The Structure Organizer: The structure organizer selects found. When it terminates, the system specifies the content of a set of additional events for suspense arousal, based on a the output sjuzhet as . heuristic function that examines the syntactical properties of CHEONG AND YOUNG: SUSPENSER: A STORY GENERATION SYSTEM FOR SUSPENSE 45 a plan structure. The function is formalized based on the fol- lowing rules assuming tha the reader is not informed of the pro- tagonist's success or failure yet: a) presenting a goal-threatening action whose effects negate the protagonist's goal/plan will increase the reader's sus- pense; b) presenting a goal-supporting action whose effects unify with the protagonist's goal/plan will decrease the reader's suspense. According to these rules, the elements of would be com- posed of a set of goal-threatening actions to invoke a higher suspense. Note that the terms goal-threatening actions and goal- supporting actions are used unconventionally in this work. To determine whether an event is goal-threatening or not, we define the term potential suspense. Heuristic Function 3 computes the potential suspense for an action by summing up the potential suspense of its effects. An event is classified as a goal-threat- ening action if its potential suspense is greater than a predefined threshold. Conversely, an event is labeled as a goal-supporting action if its potential suspense is less than a predefined value. 4) Heuristic Function 3 (Potential Suspense of an Action): returns the summation of where is the potential suspense of an effect of an action in plan .

where is the set of 's effects, is the potential sus- pense of an effect of an action in plan . Fig. 4. Plan structure extended from Fig. 1: , ,and are added to illus- trate threatening-links and supporting-links. These actions describe that the dad In computing the potential suspense of an action's effect, we is stolen his toy by a thief ( ), the police arrests the thief ( ), and the police consider all of the action's possible causal relationships to ac- returnsthetoytothedad( ). A box represents an action, with its preconditions complishing the protagonist's goal from the reader's point of on the left and effects on the right. Solid arrows denote causal links. Dashed ar- rows indicate threatening-links in which the effect at the starting point negates view. As an illustration, Fig. 4 shows that action S5:Steal has the precondition of the action at the end point. Double-lined arrows represent the effect —the negation of action S2:Put 's supporting-links where a step's effect unifies with another step's precondition. precondition (has Dad toy) —which can possibly interfere the action S2:Put with its execution from the reader's point of view where returns all the threatening-links of an effect , until the successful outcome is disclosed. We call this type of returns all the supporting-links of , and are coefficients, temporary threats which are resolved later in the story as threat- and returns a value associated with the ening-links, referring to an action's effect negating another step's causal distance between step and the goal step of plan .All precondition in the plan. In contrast, the suspense creator estab- scaling factors in the equation are constrained to be nonnegative lishes a supporting-link when the effect of an action unifies with real numbers. a precondition of another action in the plan. In the same figure, 4 S1:Trade has the effect (has Dad toy) that contains a sup- V. E VALUATION porting-link that unifies with the same condition of S2:Put.At This section describes several experiments that we carried the point where the action S1:Trade is shown the audience may out to evaluate Suspenser. An initial pilot study, described in expect that the goal of getting Ben the toy will be accomplished the first section, aimed to test a partial implementation of Sus- easily, although the later action S5:Steal will negate the (has penser and the experimental methodology. The second section Dad toy) condition. As illustrated, one effect can have multiple describes the main study we conducted. threatening-links and supporting-links in a single plan. There- fore, potential suspense of an effect is computed as the accu- A. A Pilot Study mulation of all 's supporting-links weighted by subtracted from the summation of all 's threatening-links weighted by We carried out a pilot study to evaluate the effectiveness of (Heuristic Function 4). story generation for suspense produced by a partial implementa- 5) Heuristic Function 4 (Potential Suspense of an Effect): tion of Suspenser —the skeleton builder module (i.e., Heuristic returns a value of an effect of an action in plan Function 1) and the additional event selection by Heuristic as Function 3. The reader model was not tested in this study; therefore, this pilot study did not evaluated the effectiveness of Heuristic Function 2 that measures the suspense level. The hypothesis for this study was to test if there was any association 46 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 7, NO. 1, MARCH 2015

TABLE I TABLE II EXPERIMENTAL VALUES FOR WEIGHTING CONSTANTS MEANS AND STANDARD DEVIATIONS OF SUSPENSE RATINGS

between the story generator type (independent variable) and the suspense level of the stories (response variable). To test this hypothesis, we compared the suspense levels among the stories produced by: a) Suspenser to elicit high suspense; b) a human the story. We did not constrain the number of sentences that she author; and c) a baseline that is designed to elicit low suspense selected. in the reader. 3) Procedure: We utilized a repeated measured between 1) Configuring the Experimental System: The values of the group design and assigned the subjects randomly to one of scaling factors for our heuristic functions are shown in Table I. nine subject groups. These groups were arranged according to The values of the constant factors were determined empirically a3 3 Latin Square design to counterbalance the interference from some informal experiments. In our study the value for from different orderings of stories. Each subject individually and , were initialized as 1 and 5, respectively, informed by the participated in the study by accessing a web page that presented CPI model that generates concise instructions [46]. and a sequence of three stories. Each story was divided into two were adjusted as 0.3 and 2.5, respectively. The setting of these parts: one containing the text describing the story's background coefficients placed more importance on the value of outgoing and the portion preceding the measurement point in the story, causal links than the incoming causal links, which is consistent and one containing a paragraph describing the portion of the with the views discussed in the cognitive research [47], [50]. story after the measurement point. After reading the first part on The causal-chain value of an event was assigned 2.0 for mo- a web page, the subject was asked to click the button “NEXT tivated-act ; the value of an event of other act types was as- PAGE” to proceed to the next screen in which he was asked signed 0.0. The value for the threatening-link coefficient ( ) to answer his suspense level on a seven-point Likert scale of was assigned greater than that of the supporting-link coefficient [1, 7] where 1 denotes no suspense and 7 denotes extremely ( ). These values were chosen through informal experiments suspenseful. On the completion of responding to the question, a and pilot studies such as [51] for summarization and [52] for button click led him to the next page that described the second event selection. The experimenter manually tested numerous part of the story. The subject was able to leave the survey by combinations of values and selected the one that produced the closing the survey web page anytime they wanted. most suspenseful story based on her subjective judgements. The 4) Results: A total of 25 undergraduate students ranging in optimal combination could be found by examining all possible age from 20 to 29 years old participated in this study. There were combinations and having multiple judges. The function that re- 23 males and 2 females, all recruited from a Computer Science flects an event's distance effect, in Heuristic undergraduate course at the North Carolina State University. Function 1 returned 1 and in Heuristic Func- They were given extra credit in exchange of participating in this tion 4 returned the distance from an action to the goal (i.e., the study, and were presented an alternative option. The collected minimum number of causal links that relate an action to the goal data contained 75 responses from 25 subjects, 25 responses for in a plan). each fabula. However, due to an error in reproduction of the 2) Materials: We ran Crossbow to plan three fabulas. The re- writer's selection, the responses for sjuzhets created from Fabula sulting plans consisted of 18 to 20 partially ordered steps which B were excluded from the analysis. As a result, 50 observations were manually linearized. Each plan was realized as text using a were used in this analysis. To detect a significant difference be- simple template-matching technique that mapped one plan step tween story generators and fabulas, we performed a two-way into a single sentence. We prepared a total of nine sjuzhets by ANOVA to the collected data using SAS version 9.1.3 SP4. As generating three sjuzhets for each fabula: two sjuzhets by Sus- showninTablesIIandIII,thedata indicated that the story gen- penser and one sjuzhet by a human author. To obtain human erator had no effect on the suspense level ( , generated stories, we recruited one Master's student majoring ). The story generators showed uniform performance in English at North Carolina State University, a freelance writer across the two stories. There was no statistical difference in sus- who had previously had short stories published in a local news- pense ratings for fabula ( , ). No inter- paper. She was presented with the instruction sheet followed by action effect was found between the fabula type and the story the three fabulas and their corresponding measurement point. generator type ( , ). She then was asked to select a series of sentences for each fabula 5) Discussion: The stories generated by Suspenser obtained to elicit high suspense from the reader at the specified point in the highest mean ratings (Table II shows). However, no sta- CHEONG AND YOUNG: SUSPENSER: A STORY GENERATION SYSTEM FOR SUSPENSE 47

TABLE III who chose the story events for the pilot study was asked to select ANOVA SUMMARY TABLE FOR SUSPENSE two series of sentences for each fabulas: one to arouse high sus- pense and the other to elicit low suspense form the reader when his suspense level would be measured at a given point in the story. Due to the short length of stories generated by the system, a single measurement point was selected close to the ending but where a clear outcome is not revealed yet. For instance, the measurement point for Fabula C was after reading 17th sen- tence (see Appendix A). When a series of story events preceding tistical significant difference in suspense has been found in the outcome hints the outcome, the measurement point was se- the story generator type, as the ANOVA analysis indicates lected before the those preceding events occur. The measure- (Table III). We conjecture that this result may be caused by ment point for each story was presented to the human author be- several factors described below. First, the number of subjects fore her selection. Her selection was not constrained; as a result, was too small to detect a difference, particularly with short, her two versions of a fabula differed in length within a margin similar stories. Since the skeleton serves as a base story when of 2 sentences. For fair comparisons, the length of the sjuzhet creating the stories intended for high suspense and low sus- ( ) generated by the system was set to the number of the story pense, these stories share more than 50% of the total number of events in the corresponding human author's selection for high story sentences. Therefore, we needed to find another baseline suspense up to the point where the reader's suspense was mea- stories that are distinct from the other stories. Another reason sured. The human author's selection for eliciting low suspense relatestothewayofpresentingthestorytotheparticipants.A for each fabula was used as a control narrative. The control nar- whole story was presented to the subjects at once on a single ratives differed significantly in content from the other stories. web page, which normally took them a very short time to While the stories created by Suspenser and those by the human read. Thus, the subjects did not have enough time to prepare author for arousing high suspense share 50%-80% of the total themselves to anticipate a next story event. number of story sentences, the control stories overlap with the From the lessons learned from the result of this pilot study, other stories about 20%-30% of the total number of story sen- we made some changes to the design of the main experiments. tences. One sample of the fabulas and its three sjuzhets used in First, the survey interface was modified to give the subject time this study are shown in Appendix A. to build expectation about the next story segment. One web page 3) Procedure: The study utilized a repeated measured be- shows only a single story event and the subject was required to tween group design: subjects were randomly assigned to one of click a button to proceed to read the next event. Second, the nine subject groups that were designed based on a 3 3Latin subject's suspense was measured on a five-level scale. Square. Each subject individually participated in the study by accessing a web page. He was presented with three stories and B. The Main Experiment was asked to rate his suspense level for each story at a given This section describes the experiment that we carried out to point. Each story was presented to the subject sentence by sen- evaluate the effectiveness of stories that a complete implementa- tence; one page contained only one sentence and a button click tion of Suspenser produces in terms of suspense. The hypothesis led to the next page. After reading the portion of the text pre- for our study was to test if there was any association between ceding the measurement point, the subject was asked to describe the story generator type (the independent variable) and the sus- his suspense level on a five-point scale basis ranging from no pense level (the dependent variable). To test this hypothesis, we suspense to extremely suspenseful. After responding to the ques- compared the suspense levels reported by subjects as they read tion, he was presented with the second part of the story sentence stories produced by: a) Suspenser; b) a human author intended by sentence, followed by a page asking generic questions about to create high suspense; and c) the human author intended for story coherence and enjoyment on a five-point scale ranging low suspense as a control narrative. from 1 meaning notatallto 5 meaning strongly agree. 1) Configuring the Experimental System: Table I shows 4) Data Collection: A total of 98 unpaid subjects volun- the values of the scaling factors for our heuristic functions for tarily took part in the experiment, ranging in age from 20 to this study. The function that reflects an event's distance effect, more than 50 years old (42 males, 57 females, and one no re- in Heuristic Function 1 and Heuristic Func- sponse): 72 recruited from NCSU communities including re- tion 4 returned where denotes the distance from cently graduated under/graduate students across different de- an action to the goal. The value of threshold in the algorithm partments and 26 from internet female technical communities 1 was assigned 0.07. The reader's knowledge was assumed (e.g., Systers.org). All subjects were native-speakers of English. identical to the system's plan libraries that were used to create The collected data contained 294 responses from 98 subjects. the input fabulas. The reader model's plan ranking function 5) Results: Table IV shows that the suspense ratings for preferred short plans with fewer flaws. Its reasoning resource the stories selected by Suspenser are as high as the ratings for limit was set to a search limit of 500 nodes. the stories chosen by the human author. To detect a signifi- 2) Materials: For the main experiment, the three fabulas in cant difference between three story generators, we performed a the pilot study were reused. We prepared a total of nine sjuzhets two-way ANOVA on the collected data using SAS version 9.1.3 by generating three sjuzhets for each fabula:one sjuzhet by Sus- SP4. In this analysis, two main effects were examined: the story penser and two sjuzhets by a human author. The human writer generator type and the fabula type. Each type has three levels. 48 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 7, NO. 1, MARCH 2015

TABLE IV TABLE VII MEANS AND STANDARD DEVIATIONS FOR SUSPENSE IN EACH STORY AND MEANS AND STANDARD DEVIATIONS FOR SUSPENSE,INTERESTINGNESS, AND STORY GENERATOR COHERENCE RATINGS FOR EACH STORY GENERATOR TYPE IN COHERENT NARRATIVE EXPERIENCE ONLY

TABLE VIII ANOVA SUMMARY TABLE FOR SUSPENSE IN COHERENT NARRATIVE EXPERIENCE ONLY

TABLE V ANOVA SUMMARY TABLE FOR SUSPENSE TABLE IX ONE TAILED -TEST ANALYSIS SHOWING COMPARISON OF MEANS FOR SUSPENSE.COMPARISONS SIGNIFICANT AT THE 0.01 LEVEL ARE INDICATED BY ** AND COMPARISONS SIGNIFICANT AT THE 0.05 LEVEL ARE INDICATED BY *

TABLE VI ONE-TAILED -TEST ANALYSIS SHOWING PAIR-WISE COMPARISON OF MEANS FOR SUSPENSE.COMPARISONS SIGNIFICANT AT THE 0.01 LEVEL ARE INDICATED BY ** analyzed. The data suggest that the text quality was good enough for the subjects to understand the stories. The par- ticipants evaluated the given stories as moderately coherent ( , ) on the five Likert scale of [1, 5] where 1 means not coherent and 5 means strongly coherent. We further investigated the reader's suspense for only those As shown in Table V, the data indicated that the story gener- stories that received valid coherency ratings. To this end, we ator type had an effect on the suspense level ( , eliminated the reports that contain not coherent or no response ). The result also shows that the fabula type had on the coherency question. This preprocessing step eliminated no effect on suspense. No interaction effect was found between 70 responses from the initial data, resulting in 224 responses. the fabula type and the story generator type ( , For the data analysis the R version 2.14.2 software package was ). Despite the short sample stories, the sub- used [53]. As Table VII shows, the stories created by the human jects rated their experience of suspense as moderate ( writer received the highest ratings in suspense ( ). , ) on a five-point Likert scale of [1, 5] On the other hand, Suspenser's stories are superior to human where 1 means no suspense and 5 means extremely suspenseful. author's stories in terms of interestingness ( )and A series of standard one-tailed -tests were used to compare coherence ( ). A two-way ANOVA on the data the performance of the three story generators. The results in (Table VIII) indicated that both the fabula and story generator Table VI indicate that the participants were more likely to rate types had effects on the suspense level ( ). the stories produced by Suspenser and the human author more No interaction effect was found between the fabula type and suspenseful than the control versions with a 99% of confidence the story generator type. Pearson product-moment correlation level ( , for Suspenser versus Con- analysis indicates high positive correlations among suspense, trol; , for Human versus Control). interestingness, and coherence ( ). The results of The effect sizes were small-to-moderate (Cohen's , one-tailed -tests (in Table IX) indicate that the stories pro- with a 0.05 of significance level for Suspenser duced by the system and the human author were rated as more versus Control; Cohen's , with a 0.05 suspenseful than the control stories with a 95% confidence of significance level for Human versus Control). interval and a 99% confidence interval, respectively ( , To test if the text quality affected the reader's story compre- for Suspenser versus Control; , hension, the subjects' responses to story coherency were also for Human vs. Control). The effect size CHEONG AND YOUNG: SUSPENSER: A STORY GENERATION SYSTEM FOR SUSPENSE 49

suspense ratings for Suspenser's stories and the human author's stories were different. Compared with the relatively even dis- tribution in suspense ratings for Suspenser, the ratings for the author's stories were inclined to the high-rating side. This may suggest that there can be qualitative differences between the sto- ries generated by Suspenser and those generated by the human author. On the other hand, the stories generated by the system were rated higher in interestingness and coherence than those generated by the human author. This result could be linked to the fact that the human author was asked to select story events for the effect of suspense only, whereas Suspenser considers co- herence as well by building the skeleton from the story's core story events. Therefore, further discussions regarding interest- Fig. 5. Box plot graphs of suspense ratings for coherent stories only. The axis ingness and coherence are beyond the scope of this paper that represents the story generator types and the axis represents suspense ratings. focuses on suspense only. Each box accounts for the 50% of the population and the thick horizontal bar in the box represents the median value. The top of the box denotes the upper quartile and its bottom denotes the lower quartile. Those ratings in the top 25% ofthedataareshownby the top whisker and those in the low 25% are shown VI. CONCLUSION in the bottom whisker. The dots represent outliners. Narrative generation by computers has been actively re- searched for two decades with special attention to games. Although a number of approaches have shown promise in their ability to generate narrative, there has been little research on creating stories that prompt an intended emotion. This paper presents a computational model of generating stories for sus- pense, exploring the concept that a reader's suspense level is affected by the number of solutions available to the problems faced by a narrative's protagonists [15], [16], [22], [25], [26]. When given a formal characterizationofastoryworld,this model elaborates a story content that can manipulate reader sus- pense at a specific point in its telling. Our approach gauges the suspense level that a reader would feel by modeling the reader's narrative comprehension process using a planning technique. Fig. 6. Histograms of suspense ratings for coherent story responses only. The axis represents the suspense ratings and the axis represents the count of The system takes as input a partial plan corresponding to the responses. portion of a story that has been conveyed so far and computes the reader's anticipated suspense level based on the inverse of the number of solutions to the protagonist's goals that can be was small-to-moderate for Suspenser versus Control (Cohen's found in the space of complete plans she can consider within her , with a 0.05 significance level) reasoning resources. To generate a partial plan that maximizes and moderate for Human versus Control (Cohen's , the reader's suspense, the system takes a plan as input and with a 0.05 significance level). The boxplot selects a set of core events that have high causal connectivity diagrams in Fig. 5 depict that the suspense ratings for stories and that also play an important role in the story. The partial by human author are skewed to the high ratings and the ratings plan then is supplemented by harmful actions that intensify of control stories are skewed to the low ratings. The histogram the reader's suspense level. The model has been implemented in Fig. 6 shows the counts of reports for each suspense ratings and formally evaluated. The quantitative analyses of the data grouped by the story generator. obtained from the experiments have shown this system to be 6) Discussion: The data show that the story generators had successful in selecting content that elicits high suspense. In an influence on the amount of suspense that the subjects felt. particular, the data show that, in the context of our experiments, In particular, the stories produced by Suspenser created stories this model was as effective as a human author in generating comparable in suspense to those produced by human authors. suspenseful stories. To our knowledge, Suspenser is the first The one-tailed -test results also show that the difference be- system that aims to generate suspenseful stories by modeling tween the suspense levels felt by the subjects from Suspenser's a storyteller who selects relevant story elements based on stories and the control stories was significant (a 99% confidence the reader's reasoning process. We believe that this work will interval for all stories and a 95% confidence interval for coher- benefit the AI, game, and affective computing communities. ently rated stories only). The most important limitation of the current study is the For the coherently rated story data set, the human author's sto- heuristic function that approximates the suspense level by the ries received the highest suspense ratings, resulting in the largest number of solutions for the protagonist's goal only, without effect size when compared with the stories in the control group examining the likelihood of their successful executions. De- (Cohen's ). It is also noted that the distributions of spite this limitation, Suspenser has shown successful results 50 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 7, NO. 1, MARCH 2015 in selecting a suspenseful story from a given story when com- money and kill Agatha during the charity bazaar. [3] Sykes bor- pared with a human author's performance. We believe that this rows some money from the bank by mortgaging his theater. [4] result was partly due to its subcomponents: the skeleton builder Sykes buys insurance to cover his loss in case of a fire. [5] Janet and the structure organizer. The core events identified by the gives Kent's contact information to Sykes and informs him of skeleton builder may ensure that the final sjuzhet contains Kent's expertise with firebombs. [6] Kent takes a bomb to the the suspensefulness that the input fabula is able to elicit. In Hollywood Theater and meets with Sykes. [7] Sykes purchases addition, the structure organizer selects candidate events for the firebomb. [8] Sykes installs the firebomb. [9] The lieutenant, inclusion in the sjuzhet as the events whose effects are negating Bill, issues a warrant permitting the arrest of Kent for his illegal another event's preconditions. Therefore, we conclude that the weapons dealing. [10] Bill arrests Kent. [11] Bill coaxes Kent limitation of the heuristic function in the suspense measure to give information in exchange for releasing him. [12] Kent in- component could be compensated by the other components in forms Bill that Sykes is planning to firebomb his own theater the system. A second limitation concerns the evaluation that during the charity event. [13] Bill releases Kent for his coop- uses a single human author. We recruited a person who is a eration. [14] Agatha goes to the theater for the charity event. professional writer, expecting that she would represent the [15] Sykes sets the timer of the firebomb to explode during the expert in narrative generation. However, the task of selecting charity event. [16] Sykes switches on the firebomb. [17] Bill sjuzhet from a predefined story events may leave little room to searches for the firebomb in the theater. [18] Bill defuses the express her expertise. In the current study, the writer selected firebomb. [19] Agatha participates in the charity event. 31 out of 57 events of all the fabulas, accounting for 54%. It Sjuzhet 1: Storywriter's Selection Intended for High Sus- is expected that the writer will exhibit a better performance pense: Janet and Sykes plan to burn down Sykes' theater to get when a wider selection is offered. A further evaluation can be the insurance money and kill Agatha during the charity bazaar. conducted by having fabulas consisting of a large number of Janet gives Kent's contact information to Sykes and informs him events and by recruiting multiple human authors. of Kent's expertise with firebombs. Kent takes a bomb to the Several aspects of the current system will be investigated and Hollywood Theater and meets with Sykes. Sykes purchases the extended in future work. First, we will refine the function used firebomb. Sykes installs the firebomb. Kent informs Bill that to measure the suspense level in a story. For instance, taking into Sykes is planning to firebomb his own theater during the charity account the difficulty of achieving a plan (e.g., size, readiness of event. Agatha goes to the theater for the charity event. Sykes sets executing its actions) in the human reader's reasoning process the timer of the fire-bomb to explode during the charity event. can be simulated by employing a probabilistic planning tech- Sykes switches on the firebomb. Bill searches for the firebomb nique. Second, the current approach uses a plan library that is in the theater. Bill defuses the firebomb. used for generating the input fabula plan as the reader's knowl- Sjuzhet 2: Suspenser's Selection Intended for High Sus- edge. It will be interesting and challenging to investigate the pense: Kent takes a bomb to the Hollywood Theater and meets problem of knowledge discrepancies in storytelling via the use with Sykes. Sykes purchases the firebomb. Sykes installs the of different plan libraries. firebomb. Bill arrests Kent. Kent informs Bill that Sykes is plan- ning to firebomb his own theater during the charity event. Bill releases Kent for his cooperation. Agatha goes to the theater for APPENDIX the charity event. Sykes sets the timer of the firebomb to ex- FABULA C. THE PARTS THAT WERE SHOWN AFTER SUSPENSE plode during the charity event. Sykes switches on the firebomb. MEASUREMENT WERE ITALICIZED Bill searches for the firebomb in the theater. Bill defuses the fire- Background: Sykes is the owner of the Hollywood The- bomb. Agatha participates in the charity event. ater, which was once prosperous but has now become dilapi- Sjuzhet 3: Storywriter's Selection for Low Suspense as a dated and is in need of major renovations. Sykes has accrued a Baseline Story: Janet convinces Sykes to participate in her plan sizable gambling debt, and with his theater in shambles, he has to kill Agatha by convincing him that if he participates, he will no means with which to pay it back. He is constantly threatened be able pay off his gambling debts. Sykes borrows some money by his crooked debtors. Janet is a famous actress with dreams from the bank by mortgaging his theater. Sykes buys insurance of winning an Oscar, an acting award. She is jealous of the ac- to cover his loss in case of a fire. Janet gives Kent's contact infor- tress Agatha, who is her contender for the Oscar this year and mation to Sykes and informs him of Kent's expertise with fire- also is well known for her active involvement in charity. Janet bombs. Kent takes a bomb to the Hollywood Theater and meets knows a number of scoundrels including a guy named Kent, a with Sykes. Sykes purchases the firebomb. The lieutenant, Bill, bomb dealer, and the theater owner Sykes. Agatha is in love issues a warrant permitting the arrest of Kent for his illegal with Bill, who serves as a lieutenant in the Los Angeles Police weapons dealing. Bill coaxes Kent to give information in ex- Department's Serious Crime squad. Janet knows that Agatha is change for releasing him. Bill releases Kent for his cooperation. planning to go to the Charity Bazaar for the Poor to be held in Agatha participates in the charity event. Hollywood Theater. To ensure that she will win the Oscar, Janet plans to kill Agatha during the charity event. REFERENCES The Input Fabula Story: [1] Janet convinces Sykes to par- ticipate in her plan to kill Agatha by convincing him that if he [1] Heavy Rain. Quantic Dream, 2010. [2] L.A. Noire. Team Bondi, 2011. participates, he will be able pay off his gambling debts. [2] Janet [3] BioShock Infinite. Irrational Games, 2013. and Sykes plan to burn down Sykes' theater to get the insurance [4] Fallout: New Vegas. Obsidian Entertainment, 2010. CHEONG AND YOUNG: SUSPENSER: A STORY GENERATION SYSTEM FOR SUSPENSE 51

[5] M. Cavazza, F. Charles, and S. J. Mead, “Character-based interactive [34] R. M. Young, M. E. Pollack, and J. D. Moore, “Decomposition and storytelling,” IEEE Intell.t Syst., vol. 17, pp. 17–24, Jul. 2002. Causality in Partial Order Planning,” in Proc. 2nd Int. Conf. Artif. In- [6] M. Mateas and A. Stern, “Façade: An experiment in building a fully- tell. Plan. Syst. (AIPS), 1994, pp. 188–193. realized interactive drama,” in Proc. Game Develop. Conf., 2003. [35] R. M. Young and J. D. Moore, “DPOCL: A principled approach to dis- [7] B. Mott, S. Lee, and J. Lester, “Probabilistic goal recognition in inter- course planning,” in Proc. 7th Int. Workshop Natural Lang. Generat., active narrative environments,” in Proc. AAAI, 2006. 1994, pp. 13–20. [8] R. W. Hill Jr., A. W. Hill, J. Gratch, S. Marsella, J. Rickel, W. [36] D. Pizzi, F. Charles, J. luc Lugrin, and M. Cavazza, “Interactive story- Swartout, and D. Traum, “Virtual humans in the mission rehearsal telling with literary feelings,” in Proc. 2nd Int. Conf. Affect. Comput. exercise system,” KI Embodied Conversat. Agents, vol. 17, pp. 32–38, Intell. Interaction (ACII2007), 2007. 2003. [37] F. Peinado, M. Cavazza, and D. Pizzi, “Revisiting character-based af- [9] N. Szilas, “A new approach to interactive drama: From intelligent char- fective storytelling under a narrative BDI framework,” in LNCS 5334: acters to an intelligent virtual narrator,” in Proc. AAAI Spring Symp. AI Interactive Storytelling, U. Spierling and N. Szilas, Eds. Berlin, : Interact. Entertain., 2001, pp. 72–76. Springer-Verlag, 2008, pp. 83–88. [10] W. F. Brewer and E. H. Lichtenstein, “Stories are to entertain: A struc- [38] J. Platts, A. Blandford, and C. Huyck, “Awaiting the sensation of a tural-affect theory of stories,” J. Pragmat., vol. 6, no. 5-6, pp. 473–486, short, sharp shock: Twist-centred story generation by transformation,” Dec. 1982. in Proc. AISB'02 Symp. AI Creativity Arts Sci., 2002. [11] W. F. Brewer and K. Ohtsuka, “Story structure, characterization, just [39] B.-C. Bae and R. M. Young, “A use of flashback and foreshadowing world organization, and reader affect in american and hungarian short for surprise arousal in narrative using a plan-based approach,” in stories,” Poetics, vol. 17, no. 4-5, pp. 395–415, 1988. LNCS 5334: Interactive Storytelling,U.SpierlingandN.Szilas, [12] L. F. Alwitt, “Suspense and advertising responses,” J. Consumer Psy- Eds. Berlin, Germany: Springer-Verlag, 2008, pp. 156–167. chol., vol. 12, no. 1, pp. 35–49, 2002. [40] C. B. Callaway and J. C. Lester, “Narrative prose generation,” Artif. [13] A. Ortony, G. L. Clore, and A. Collins, The Cognitive Structure of Emo- Intell., vol. 139, pp. 213–252, Aug. 2002. tions. Cambridge, U.K.: Cambridge Univ. Press, 1988. [41] A. Jhala and R. M. Young, “Cinematic visual discourse: Representa- [14] P. Vorderer, H. J. Wulff, and F. Mike, “Toward a psychological theory tion, generation, and evaluation,” IEEE Trans. Computat. Intell. AI in of suspense,” in Suspense: Conceptualizations, Theoretical Analyses, Games, vol. 2, no. 2, pp. 69–81, Jun. 2010. and Empirical Explorations. Mahwah, NJ, USA: Lawrence Erlbaum [42] R. Barthes and L. Duisit, “An introduction to the structural analysis of Associates, 1996. narrative,” New Literary History, vol. 6, no. 2, pp. 237–272, 1975. [15] D. Zillmann, “The psychology of suspense in dramatic exposition,” [43] S. B. Chatman, Story and Discourse: Narrative Structure in Fiction in Suspense: Conceptualizations, Theoretical Analyses, and Empirical and Film. Ithaca, NY, USA: Cornell Univ. Press, 1980. Explorations,P.Vorderer,H.J.Wulff,andM.Friedrichsen,Eds. [44] M. J. Rattermann, L. Spector, J. Grafman, H. Levin, and H. Har- Mahwah, NJ: Erlbaum, 1996, pp. 199–232. ward, “Partial and total-order planning: Evidence from normal and [16] R. J. Gerrig and A. B. I. Bernardo, “Readers as problem-solvers in the prefrontally damaged populations,” Cogn. Sci.,vol.25,no.6,pp. experience of suspense,” Poetics, vol. 22, no. 6, pp. 459–472, 1994. 941–975, 2001. [17] G. Genette, Narrative Discourse. Oxford, U.K.: Basil Blackwell, [45] S. Kambhampati, C. A. Knoblock, and Q. Yang, “Planning as refine- 1980. ment search: A unified framework for evaluating design tradeoffs in [18] M. Bal, Narratology: Introduction to the Theory of Narra- partial-order planning,” Artif. Intell., vol. 76, no. 1-2, pp. 167–238, tive. Toronto, ON, Canada: University of Toronto Press, 1985. 1995. [19] S. Rimmon-Kenan, Narrative Fiction: Contemporary Poetics,2nd [46] R. M. Young, “Using Grice's maxim of quantity to select the content ed. London, U.K.: Routledge, 2002, New accents. of plan descriptions,” Artif. Intell., vol. 115, pp. 215–256, Dec. 1999. [20] M. J. Toolan, Narrative: A Critical Linguistic Introduction,2nded. [47] T. Trabasso and L. L. Sperry, “Causal relatedness and importance of London, U.K.: Routledge, 2001, Interface. story events,” J. Memory Lang., vol. 24, no. 5, pp. 595–611, 1985. [21] D. Zillmann, “The logic of suspense and mystery,” in Responding to the [48] C. L. Foss and G. H. Bower, “Understanding actions in relation to Screen: Reception and Reaction Processes, Bryant and D. Zillmann, goals,” in Advances in Cognitive Science,N.E.Sharkey,Ed. Chich- Eds. Hillsdale, NJ, USA: Lawrence Erlbaum, 1991, pp. 281–303. ester, U.K.: Ellis Horwood, 1986, vol. I, pp. 94–124. [22] W. F. Brewer, “The nature of narrative suspense and the problem of [49] W. van der Hoek and A. Lomuscio, “A logic for ignorance,” in LNCS rereading,” in Suspense: Conceptualizations, Theoretical Analyses, 2990: Declarative Agent Languages and Technologies. Berlin, Ger- and Empirical Explorations, P. Vorderer, H. J. Wulff, and M. many: Springer-Verlag, 2004, pp. 97–108. Friedrichsen, Eds. Mahwah, NJ, USA: Lawrence Erlbaum Asso- [50] A. C. Graesser, K. L. Lang, and R. M. Roberts, “Question answering ciates, 1996, pp. 107–127. in the context of stories,” J. Experiment. Psychol.: General, vol. 120, [23] R. J. Gerrig, “The resiliency of suspense,” in Suspense: Concep- no. 3, pp. 254–277, 1991. tualizations, Theoretical Analyses, and Empirical Explorations,P. [51] Y.-G. Cheong and R. M. Young, “A computational model of narrative Vorderer, H. J. Wulff, and M. Friedrichsen, Eds. Mahwah, N.J., generation for suspense,” in Proc. 21st AAAI 2006 Computat. Aesthetic USA: Lawrence Erlbaum Associates, 1996, pp. 93–105. Workshop, 2006, pp. 1906–1907. [24] P. Comisky and J. Bryant, “Factors involved in generating suspense,” [52] Y.-G. Cheong, A. Jhala, B.-C. Bae, and R. M. Young, “Automatically Human Commun. Res., vol. 9, no. 1, pp. 49–58, 1982. generating summary visualizations from game logs,” in Proc. AIIDE, [25] N. Carroll, “The paradox of suspense,” in Suspense: Conceptualiza- 2008. tions, Theoretical Analyses, and Empirical Explorations, P. Vorderer, [53] R. D. C. Team, R: A Language and Environment for Statistical Com- H. J. Wulff, and M. Friedrichsen, Eds. Mahwah, NJ, USA: Lawrence puting. Vienna, Austria: R Foundation, 2012, Statistical Computing. Erlbaum Associates, 1996, pp. 71–92. [26] M. de Wied, “The role of temporal expectancies in the production of film suspense,” Poetics, vol. 23, no. 1-2, pp. 107–123, 1995. Yun-Gyung Cheong (M’10) received the B.S. de- [27] N. M. Sgouros, “Dynamic generation, management and resolution of gree, in 1996, and the M.S. degree, in 1998, in infor- interactive plots,” Artif. Intellig., vol. 107, pp. 29–62, Jan. 1999. mation engineering from Sungkyunkwan University, [28] M. Lebowitz, “Story-telling as planning and learning,” Poetics,vol.14, South Korea. In 2007, she received the Ph.D. degree no.6,1985. in computer science from North Carolina State Uni- [29] B. Magerko and J. E. Laird, Mediating the Tension Between Plot and versity, Chapel Hill, NC, USA. Interaction. San Jose, CA: , 2004, 2004 AAAI Workshop Series: She worked as a Researcher at Samsung Advanced Challenges in Game Artificial Intelligence. Institute of Technology, South Korea (2008–2010), [30] R. Aylett, “Narrative in virtual environments-towards emergent narra- and was a Postdoctoral Fellow of the Center for Com- tive,” in Proc. Narrative Intell. Symp., 1999. puter Games Research, IT University of Copenhagen, [31] S. R. Turner, The Creative Process: A Computer Model of Storytelling in Denmark (2010–2014). She is currently a Research and Creativity. Hillsdale, NJ, USA: Lawrence Erlbaum Assoc., 1994. Professor at the College of ICE at Sungkyunkwan University. Her research in- [32] M. O. Riedl and R. M. Young, “Narrative planning: Balancing plot and terests include artificial intelligence with emphasis on its use in narrative plan- character,” J. Artif. Intell. Res., vol. 39, 2010. ning, games, and data mining. [33] S. G. Ware and R. M. Young, “CPOCL: A narrative planner supporting Dr. Cheong is a Member of the IEEE Games Technical Committee (GTC) of conflict,” in Proc. AIIDE,2011. the Computational Intelligence Society. 52 IEEE TRANSACTIONS ON COMPUTATIONAL INTELLIGENCE AND AI IN GAMES, VOL. 7, NO. 1, MARCH 2015

R. Michael Young (M’03–SM’04) received the is currently a Professor of Computer Science and University Faculty Scholar B.S. degree in computer science, in 1984, from the at North Carolina State University, Raleigh, NC, USA, where he directs the California State University at Sacramento, Sacra- Digital Games Research Center. His research interests span the computational mento, CA, USA, and the M.S. degree in computer modeling of narrative, the use of artificial intelligence techniques within science with a concentration in symbolic systems computer games and virtual worlds. from Stanford University, Stanford, CA, USA, in Dr. Young is a Senior Member of the Association for the Advancement of 1988. In 1998, he received the Ph.D. degree in Artificial Intelligence and is an Association for Computing Machinery Distin- intelligent systems from the University of Pittsburgh, guished Scientist. Pittsburgh, PA, USA. He worked as a Postdoctoral Fellow at the Robotics Institute at Carnegie Mellon University. He